Learnings

---
name: learnings
description: "Final pre-commit sweep — review the session and code changes for anything worth codifying in CLAUDE.md or README.md."
---

# Learnings

Final pre-commit sweep: did we learn anything in this session that the next person (or agent) working in this repo will regret not having written down?

The bar is **high**. Most sessions produce nothing worth adding. That's fine. Saying "nothing worth adding" is a valid — and often correct — outcome. Do not invent findings to justify the skill.

## Step 1: Gather context

Run these in parallel:
- `git diff` (unstaged) and `git diff --cached` (staged) — what actually changed
- `git status` — new files, deleted files
- `git log --oneline -20` — recent commit style and scope
- Read `CLAUDE.md` if it exists (project root, then `.claude/CLAUDE.md`)
- Read `README.md` if it exists

Also re-read this session's conversation: what did the user correct you on? What did you stumble over? What non-obvious thing did you discover about the codebase?

## Step 2: Apply the bar

A candidate learning is worth codifying **only** if it meets at least one of these:

1. **Non-obvious gotcha** — something that cost real time this